Transformer for classification

Abstract

A method may embed the input dataset into a first embedding space. A method may input the first embedding space into a spectral module including a periodic information processor and an aperiodic information processor. A method may identify global features in the input dataset using the periodic information processor based on a first subset of the first embedding space. A method may identify first local features in the input dataset using the aperiodic information processor based on a second subset of the first embedding space, wherein the first subset and the second subset are different. A method may combine the global features and the first local features into a dataset of classified features of the input dataset.

What is claimed is: 
     
         1 . A method of classifying an input dataset, the method comprising:
 embedding the input dataset into a first embedding space;   inputting the first embedding space into a spectral module including a periodic information processor and an aperiodic information processor;   identifying global features in the input dataset using the periodic information processor based on a first subset of the first embedding space;   identifying first local features in the input dataset using the aperiodic information processor based on a second subset of the first embedding space, wherein the first subset and the second subset are different; and   combining the global features and the first local features into a dataset of classified features of the input dataset.   
     
     
         2 . The method of  claim 1 , further comprising:
 identifying second local features in the input dataset using an attention processor based on output of the spectral module; and   combining the second local features with the global features and the first local features in the dataset of the classified features of the input dataset.   
     
     
         3 . The method of  claim 1 , wherein the periodic information processor identifies global features in the input dataset using a frequency domain. 
     
     
         4 . The method of  claim 3 , wherein the periodic information processor translates the first embedding space from a spatial domain into the frequency domain using a Hartley transformation, identifies at least one of the global features using a spectral gating network, and applies an inverse Hartley transformation to translate the first embedding space from the frequency domain to the spatial domain. 
     
     
         5 . The method of  claim 1 , wherein the aperiodic information processor identifies local features in the input dataset using at least one Hartley convolutional transformer. 
     
     
         6 . The method of  claim 1 , wherein the first subset and the second subset are mutually exclusive. 
     
     
         7 . The method of  claim 1 , wherein the first subset and the second subset combine to yield the input dataset. 
     
     
         8 . A computing system for classifying an input dataset, the computing system comprising:
 one or more hardware processors;   an embedding processor executable by one or more hardware processors and configured to embed the input dataset into a first embedding space;   a spectral processor executable by the one or more hardware processors and configured to input the first embedding space into a spectral module including a periodic information processor and an aperiodic information processor, wherein the periodic information processor is configured to identify global features in the input dataset based on a first subset of the first embedding space and the aperiodic information processor is configured to identify first local features in the input dataset based on a second subset of the first embedding space, wherein the first subset and the second subset are different; and   an output interface executable by the one or more hardware processors and configure to combine the global features and the first local features into a dataset of classified features of the input dataset.   
     
     
         9 . The computing system of  claim 8 , wherein the spectral processor outputs a second embedding space and further comprising:
 an attention processor executable by the one or more hardware processors and configured to identify second local features in the input dataset using attention processing based on the second embedding space, wherein the output interface is further configured to combine the second local features with the global features and the first local features in the dataset of the classified features of the input dataset.   
     
     
         10 . The computing system of  claim 8 , wherein the periodic information processor identifies global features in the input dataset using a frequency domain. 
     
     
         11 . The computing system of  claim 10 , wherein the periodic information processor is configured to translate the first embedding space from a spatial domain into the frequency domain using a neural operator, identify at least one of the global features using a spectral gating network, and apply an inverse transformation to translate the first embedding space from the frequency domain to the spatial domain. 
     
     
         12 . The computing system of  claim 8 , wherein the aperiodic information processor identifies local features in the input dataset using at least one convolutional operator. 
     
     
         13 . The computing system of  claim 8 , wherein the input dataset is an image. 
     
     
         14 . The computing system of  claim 8 , wherein the first subset and the second subset combine to yield the input dataset. 
     
     
         15 . One or more tangible processor-readable storage media embodied with instructions for executing on one or more processors and circuits of a computing device a process for classifying an input dataset, the process comprising:
 embedding the input dataset into a first embedding space;   inputting the first embedding space into a spectral module including a periodic information processor and an aperiodic information processor;   identifying global features in the input dataset using the periodic information processor based on a first subset of the first embedding space;   identifying first local features in the input dataset using the aperiodic information processor based on a second subset of the first embedding space; and   combining the global features and the first local features into a dataset of classified features of the input dataset.   
     
     
         16 . The one or more tangible processor-readable storage media of  claim 15 , wherein the process further comprises:
 identifying second local features in the input dataset using an attention processor based on output of the spectral module; and   combining the second local features with the global features and the first local features in the dataset of the classified features of the input dataset.   
     
     
         17 . The one or more tangible processor-readable storage media of  claim 15 , wherein the periodic information processor identifies global features in the input dataset using a frequency domain. 
     
     
         18 . The one or more tangible processor-readable storage media of  claim 17 , wherein the periodic information processor translates the first embedding space from a spatial domain into the frequency domain using a Hartley Transformation, identifies at least one of the global features using a spectral gating network, and applies an Inverse Hartley Transformation to translate the first embedding space from the frequency domain to the spatial domain. 
     
     
         19 . The one or more tangible processor-readable storage media of  claim 15 , wherein the aperiodic information processor identifies local features in the input dataset using at least one convolutional operator. 
     
     
         20 . The one or more tangible processor-readable storage media of  claim 15 , wherein the first subset and the second subset are mutually exclusive and combine to yield the input dataset.